Mammogram Classification for Malignancy Localizaton Using Attention Learning

Abstract

Mammography is an x-ray imaging method used for breast examination.. It is a specially designed medical imaging procedure to detect breast cancer at their early stage. A mammography exam report, called the mammogram assists in early identification and diagnosis of breast cancer. This project intends to classify the mammography test scan images into their corresponding classes and uses a model which combines both CNN and RNN, called attention learning to locate specific pixels of tumour using a heat map overlay. Attention learning classifier model is a classic encoder-decoder circuit where convolutional neural networks carry out encoding and recurrent neural networks carry out decoding. Convolutional neural extract features from the mammography scans which is then fed into a recurrent neural network that concentrates the malignancy region based on continuously adjusting the weights by receiving feedback over several iterations. Mammography images are normalized, enhanced and augmented before feature extraction and assign weights to them as a part of pre-processing steps. This procedure would most importantly assist in tumour localization in the case of breast cancer.